Download preference based on reasons

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Quantum logic wikipedia , lookup

Intuitionistic logic wikipedia , lookup

List of first-order theories wikipedia , lookup

Mathematical logic wikipedia , lookup

History of the function concept wikipedia , lookup

Law of thought wikipedia , lookup

Arrow's impossibility theorem wikipedia , lookup

Structure (mathematical logic) wikipedia , lookup

Propositional calculus wikipedia , lookup

Modal logic wikipedia , lookup

Interpretation (logic) wikipedia , lookup

Model theory wikipedia , lookup

Laws of Form wikipedia , lookup

Principia Mathematica wikipedia , lookup

Transcript
T HE R EVIEW OF S YMBOLIC L OGIC
Volume 5, Number 1, March 2012
PREFERENCE BASED ON REASONS
DANIEL OSHERSON
Department of Psychology, Princeton University
SCOTT WEINSTEIN
Department of Philosophy, University of Pennsylvania
Abstract. We describe a logic of preference in which modal connectives reflect reasons to desire
that a sentence be true. Various conditions on models are introduced and analyzed.
§1. Introduction. Sometimes preferences are the result of identifiable reasons, as
when you install a fire alarm for concern about safety. This suggests that studying reasons
along with preference might illuminate both. An obstacle to such a project is the multitude
of reasons for and against a given action that combine subliminally to yield a decision,
as when you go ahead with the fire alarm despite the bother, cost, and false alerts; the
underlying calculus of reason aggregation seems largely hidden from introspection. The
centrality of reasons to action and rationality is nonetheless sufficient motive to persevere
in their analysis despite the difficulty. In this spirit, we here advance a modal logic in which
different reasons for a preference can be aggregated in various ways.
Our inquiry is preceded by several studies of the logic of preference, beginning with
von Wright (1963). More contemporary work includes systems designed to elucidate the
interaction between choice and epistemic possibility (see Lang et al., 2003; van Benthem
et al., 2009). Of particular relevance is Liu (2008, chap. 3). This work introduces “priorities” (which function like reasons in the present setting) that are ordered by importance,
and integrated into a formal language of preference and belief. Several ways of extracting
preferences from priorities are explored. The interplay of preferences and beliefs is also
analyzed, along with the impact of updating belief and preference. Liu’s work is closest
to the approach taken here inasmuch as it develops a modal language and associated
semantics. The conceptual framework is nonetheless different from ours, as will become
clear presently. Another fruitful perspective on the integration of preferences issues from
the graph-theoretic approach advanced in Andréka et al. (2002); different graphs represent alternative orderings of the alternatives in play, and might be considered separate
reasons for choice among them. Within a yet different tradition, multiattribute utility theory
(Keeney & Raiffa, 1993) bears directly on reason aggregation through the combination of
utilities based on separate dimensions. The theory has revealed exact conditions under
which aggregation can proceed additively but it does not explore the logical structure of
reasons and preference, as we shall do here. The insight achievable by formal analysis of
reasons is illustrated by Dietrich & List (2009). These authors demonstrate a representation
theorem relating choice to the respective bundles of reasons that apply to the options
in play; the axioms needed for their result are remarkably weak. Several issues are thereby
Received: September 17, 2011.
122
c Association for Symbolic Logic, 2011
doi:10.1017/S1755020311000244
PREFERENCE BASED ON REASONS
123
clarified, among them the significance of combining reasons (their analysis rests not on
individual reasons but on sets of them).
To keep the present project manageable, conceptual issues about the nature of reasons
and their role in rational discourse will be set aside. An entry to this literature is provided
by Dietrich & List (2009), and sustained discussion is available in Pettit (2002). Of course,
the reasons that come to mind are not necessarily those that govern choice (see, e.g.,
Messick, 1985; Haidt, 2001). Our theory is indifferent to this distinction but it will be
more natural to limit examples to conscious effective reasons. The case of the fire alarm
serves to convey the character of our theory. Specifically, we picture an agent who imagines
a world that resembles the actual one but with a fire alarm, and another world (possibly
his own) without one. The agent then compares the two worlds according to various utility
scales (one that measures safety, another cost, and so forth), as well as a distinct utility
scale that takes all the individual scales into account. Our formalism is designed to capture
this picture.
We proceed by first introducing the language under investigation. Informal glosses for
some of its formulas will clarify the ideas in play. Next the semantics of our logic is
presented, followed by consideration of subclasses of models that meet various conditions.
We then turn to decidability issues. A discussion of open questions is provided at the end.
§2. Language. The present section introduces a family of modal languages, and discusses the intended meaning of the modality. A language of reason-based preference is
determined by its signature, which consists of:
(a) a nonempty set P of propositional variables
(b) a nonempty collection S of nonempty subsets of N (the set {0, 1, . . . } of natural
numbers)
The language of reason-based preference determined by signature (P, S) is denoted
L(P, S), and is built from the following symbols.
(a)
(b)
(c)
(d)
(e)
the set P of propositional variables
the unary connective ¬
the binary connective ∧
for every set X ∈ S, the binary connective X
the two parentheses
Formulas are defined inductively via:
p ∈ P | ¬ϕ | (ϕ ∧ ψ) | (ϕ X ψ) for X ∈ S.
Moreover, we rely on the following abbreviations.
(ϕ ∨ ψ)
for
¬(¬ϕ ∧ ¬ψ)
(ϕ → ψ)
for
(¬ϕ ∨ ψ)
(ϕ ↔ ψ)
for
((ϕ → ψ) ∧ (ψ → ϕ))
(ϕ 1...k ψ)
for
(ϕ {1...k} ψ)
(ϕ X ψ)
for
(ϕ X ψ) ∧ ¬(ψ X ϕ)
(ϕ ≈ X ψ)
for
(ϕ X ψ) ∧ (ψ X ϕ)
124
DANIEL OSHERSON AND SCOTT WEINSTEIN
(ϕ X ψ)
for
(ψ X ϕ)
(ϕ ≺ X ψ)
for
(ψ X ϕ)
for
( p → p)
⊥
for
¬
The formula ϕ 1 ψ is to be understood along the following lines. Fix an agent A whose
reasoning is at issue. Let u 1 be a utility scale that reflects some dimension of interest to A.
Then ϕ 1 ψ is true just in case:
A envisions a situation in which ϕ is true and that otherwise differs little
from his actual situation (if ϕ is already true then A’s actual situation
may well be the one he envisions). Likewise, A envisions a second situation that is like his actual situation except that ψ is true. Finally, the
utility according to u 1 of the first imagined situation exceeds that of the
second.
In the fire alarm example, A envisions his home with a new fire alarm, but with the same
furniture, cat, and fireplace as before. Home with no fire alarm is the actual situation, hence
especially easy to envision. If u 1 measures safety, and p is “A will purchase a fire alarm”
then p 1 ¬ p holds inasmuch as the alarm improves safety. (Since is also true in A’s
situation, p 1 ¬ p is materially equivalent to p 1 .) If A is short on cash, and u 2
reflects finances then p ≺2 ¬ p is true, whereas the status of p 1,2 ¬ p depends on the
manner in which utilities are aggregated (e.g., averaging, minimum, etc.). More generally,
we allow preferences ϕ X ψ between arbitrary formulas ϕ, ψ in view of the (possibly
multiple) utilities in X ∈ S. The formula ϕ X ψ thus represents
A’s preference for ϕ
over ψ when A brings to mind just the reasons indexed in X . If S ∈ S then preference
tout court for ϕ over ψ is represented by ϕ S ψ, that is, taking account of all reasons
in play.
Of course, the greater utility of a given situation compared to another is just one way of
expressing a reason for preferring the former to the latter. More generality can be achieved
by representing each kind of reason by an arbitrary binary relation over situations, instead
of insisting on numerical comparisons of cardinal utility. Recourse to such relations will
be raised again in the Discussion section. For now, we develop our theory in the context of
utility, with the expectation that most readers will find this setting conceptually familiar.
If our agent is presumed to be moral then reasons are meant, very roughly, to be good (at
least, not bad ). Morality will here be left unexplored, however. Instead, A is conceived as
logically empowered but otherwise like the rest of us. Also notice how little any of this has
to do with reasons to believe (except for odd cases like being rewarded for reaching genuine
religious conviction). Only reasons for preference will be at issue. There is nonetheless one
connection to belief that bears comment.
The appeal to situations that differ minimally from the actual one, except for satisfying a given formula, is familiar from well-known theories of counterfactual conditionals
(Stalnaker, 1968; Lewis, 1973). It thus risks bedevilment from a similar range of cases.
Suppose, for example, that p is “Winter ends a little earlier than last year.” Then too
many p-worlds offer themselves as alternatives to the actual world (since the set of shorter
winters has no member closest to last year’s winter). The present endeavor, however, may
not be as vulnerable as the earlier one to such cases. For it here suffices that the reasoning
agent bring to mind a cognitively salient situation that satisfies the formula in question
(e.g., winter a week shorter), not necessarily the maximally similar one. Indeed, the agent
PREFERENCE BASED ON REASONS
125
may not be prepared to identify the maximally similar p-world, or even to understand
such an idea. Consistent with this relaxed attitude, to each consistent proposition our
semantics assigns a world that represents life were the proposition true, where the choice
of world may depend on the agent’s current position. Some constraints on the choice will
be examined, but otherwise the reasoning agent is on his own. We take all this to be a
rough idealization of what happens in actual decision making. One imagines an alternative
situation that satisfies the proposition at issue, then evaluates it along various dimensions
(i.e., utility scales).
The utility scales that determine the truth of modal formulas are intended to measure
the impact on choice of specific considerations, for example, cost, health, professional
advancement. Because deliberation is assumed to transpire in a single mind (the agent’s),
aggregation of different scales into an overall value seems feasible; indeed, people do it
all the time. For simplicity, the scales express expected utilities, that is, with probabilities
already factored in. Thus, the safety improvements envisioned from installing a fire alarm
already integrate the agent’s confidence that the device will work as advertised.
Even when utility scales are kept separate, languages of reason-based preference allow
interesting interactions. For an illustration, first observe that ϕ i means (roughly) that
the u i -utility of the envisioned ϕ-world exceeds that of the actual world. Now consider:
( p 1 ) 2 .
This says that the agent has a u 2 -reason for there being a u 1 -reason in favor of p. For
example, let p be the assertion that you buy a low-power automobile. Let u 2 -utility be
pecuniary: u 2 (w1 ) > u 2 (w2 ) iff you have more cash in w1 compared to w2 . Let u 1 -utility
reflect personal safety: u 1 (w1 ) > u 1 (w2 ) iff you incur less risk traveling in w1 than in w2 .
Then the formula asserts that it’s in your financial interest that your buying a low-power
automobile is in your safety interest—which might well be true inasmuch as low-power
vehicles are cheaper.
We conclude this section with another illustration of the interaction of individual utility
scales. Consider:
¬q 1 ( p 2 q).
This says that the agent u 1 -prefers that q be false rather than u 2 -prefer p over q. For
example, let q be the assertion that your brother runs for mayor, and let p be that Miss
Smith (no relation) also runs. Let u 1 -utility measure family pride, and let u 2 -utility measure
political value to an ailing municipality. Then the formula asserts that from the point of
view of family pride, you’d rather that your brother not run for mayor than that Miss Smith
be the superior candidate.
§3. Semantics. We now provide a formal semantics designed to capture the intuitive picture elaborated in the preceding section. Several preliminary concepts are needed.
Fundamental is the choice of a nonempty set W to embody the imaginative possibilities
(“worlds”) available to an agent in the course of practical deliberation. Subsets of W are
called propositions. As discussed above, given a nonempty proposition A and a world w,
an agent envisions a salient alternative to w among the worlds in A. (If w ∈ A then the
“alternative” might be w itself.) We formalize this idea as follows.
D EFINITION 3.1. A selection function s over W is a mapping from W × {A ⊆ W | A = ∅}
to W such that for all w ∈ W and ∅ = A ⊆ W, s(w, A) ∈ A.
126
DANIEL OSHERSON AND SCOTT WEINSTEIN
Thus, s(w, A) is a choice of world to represent A, where the choice depends on w. (The
idea is that s chooses a member of A that is similar to w.)
Next, recall that each world can be evaluated according to various utility scales, each
involving one or more dimensions of value. All the scales are indexed by members of S.
D EFINITION 3.2. A utility function u over W and S is a mapping from W × S to (the
reals).
For w ∈ W and {i}, X ∈ S, we write u(w, {i}) as u i (w), and u(w, X ) as u X (w).
Let P be a nonempty set of propositional variables. Our last preliminary is the assignment of a proposition to each variable in P.
D EFINITION 3.3. A truth-assignment (over W and P) is a mapping from P to the power
set of W.
For a truth-assignment t, the idea is that p ∈ P is true in w ∈ W just in case w ∈ t ( p) (and
otherwise false). This is all we need to introduce models.
D EFINITION 3.4. A model for a signature (P, S) is a quadruple (W, s, u, t) where
(a)
(b)
(c)
(d)
W is a nonempty set of worlds;
s is a selection function over W;
u is a utility function over W and S;
t is a truth-assignment over W and P.
It remains to specify the proposition (set of worlds) expressed by a formula ϕ in a model
M. This proposition is denoted ϕ[ M ], and defined inductively as follows.
D EFINITION 3.5. Let signature (P, S), ϕ ∈ L(P, S), and model M = (W, s, u, t) for
(P, S) be given.
(a)
(b)
(c)
(d)
If ϕ ∈ P then ϕ[ M ] = t (ϕ).
If ϕ is the negation ¬θ then ϕ[ M ] = W \ θ [ M ].
If ϕ is the conjunction (θ ∧ ψ) then ϕ[ M ] = θ [ M ] ∩ ψ[ M ].
If ϕ has the form (θ X ψ) for X ∈ S, then ϕ[ M ] = ∅ if either θ [ M ] = ∅ or
ψ[ M ] = ∅. Otherwise:
ϕ[ M ] = {w ∈ W | u X (s(w, θ [ M ])) ≥ u X (s(w, ψ[ M ]))}.
Observe that (θ X ψ)[ M ] is defined to be empty if there is no world that satisfies θ
or none that satisfies ψ. Thus, we read (θ X ψ) with existential import (“the θ -world
is weakly X -better than the ψ-world,” where the definite description is Russellian). In the
nontrivial case, let A = ∅ be the proposition expressed by θ in M, and B = ∅ the one
expressed by ψ. Then (intuitively) world w satisfies (θ X ψ) in M iff the world selected
from A as closest to w has utility no less than that of the world selected from B as closest
to w. A word of caution: the existential requirement on the truth of (θ X ψ) allows
¬(θ X ψ)[ M ] = (θ ≺ X ψ)[ M ]. Indeed, if θ [ M ] = ∅ then ¬(θ X ψ)[ M ] = W
but (θ ≺ X ψ)[ M ] = ∅.
The following definition imports standard terminology and notation to the present context.
D EFINITION 3.6. Let ϕ ∈ L(P, S) and model M = (W, s, u, t) for (P, S) be given.
(a) M satisfies ϕ just in case ϕ[ M ] = ∅.
PREFERENCE BASED ON REASONS
127
(b) ϕ is valid in M just in case ϕ[ M ] = W.
(c) ϕ is valid just in case ϕ is valid in every model.
(d) ϕ is valid in a given class C of models just in case ϕ is valid in every model of C.
We use related expressions (like “satisfiable”) in the obvious way. It is noteworthy that our
language allows expression of the global modality (see Blackburn et al., 2001, sec. 2.1).
Choose any X ∈ S, and for ϕ ∈ L(P, S) let:
def
ϕ = ¬(¬ϕ X ¬ϕ)
and
def
♦ϕ = (ϕ X ϕ).
(3.7)
Then unwinding clause Definition 3.5d of our semantic definition yields:
P ROPOSITION 3.8. For all ϕ ∈ L(P, S) and models M = (W, s, u, t):
(a) ϕ[ M ] = ∅ iff ϕ[ M ] = W iff ϕ[ M ] = W.
(b) ♦ϕ[ M ] = ∅ iff ♦ϕ[ M ] = W iff ϕ[ M ] = ∅.
It follows from Proposition 3.8 that the axioms of S5 are valid for and ♦. Other valid
formulas of our language include the following (proofs are easy). For all X ∈ S, and
ϕ, ψ, θ ∈ L(P, S):
((ϕ X ψ) ∧ (ψ X θ)) → (ϕ X θ)
(♦ϕ ∧ ♦ψ) → ((ϕ X ψ) ∨ (ψ X ϕ))
(♦ϕ ∧ ♦ψ) ↔ (¬(ϕ X ψ) ↔ (ψ X ϕ))
¬(⊥ X ϕ) and ¬(ϕ X ⊥)
♦ϕ → (ϕ ≈ X ψ) if ϕ and ψ are equivalent.
The next two sections introduce subclasses of structures which conform to various hypotheses; we explore the logical principles validated thereby. The hypotheses considered in
Section §4 are called frame properties because their definition depends on just the worlds
and selection function of a model, that is, on just its “frame.” The remaining hypotheses
(Section §5) involve utility and the assignment of propositions to variables. Several of
the properties discussed below have already appeared within order-theoretic approaches to
preference, for example, in Levi (1986, chap. 6).
§4. Stronger theories based on frame properties of models. For this section, let
model M = (W, s, u, t) have signature (P, S).
4.1. Reflexivity. If a world w satisfies a formula ϕ then the “nearest” ϕ-world is intuitively w itself. This condition is not imposed on selection functions by Definition 3.1 but
can be added as follows.
D EFINITION 4.1. M is reflexive just in case for all w ∈ W and A ⊆ W, if w ∈ A, then
s(w, A) = w.
The formula exhibited in the following proposition illustrates the impact of reflexivity.
It says that a given proposition is at least as good as the status quo or its negation is.
P ROPOSITION 4.2. Let ϕ be ( p X ) ∨ (¬ p X ). Then ϕ is invalid but valid in the
class of reflexive models.
128
DANIEL OSHERSON AND SCOTT WEINSTEIN
Proof. To verify the invalidity of ϕ, suppose that W = {w0 , w1 , w2 }, t ( p) = {w0 , w1 },
s(w0 , {w0 , w1 }) = s(w0 , p[ M ]) = w1 , s(w0 , {w2 }) = s(w0 , ¬ p[ M ]) = w2 ,
s(w0 , W) = s(w0 , [ M ]) = w0 , and u X (w0 ) > u X (w1 ), u X (w2 ). Then it is easy to
see that w0 ∈ ϕ[ M ] hence ϕ is not valid.
On the other hand, suppose that M is reflexive, and let w0 ∈ W. Then either w0 ∈
p[ M ] or w0 ∈ ¬ p[ M ], say the former (the other case is parallel). By reflexivity,
s(w0 , p[ M ] = w0 . Likewise, w0 ∈ [ M ] = W, so again by reflexivity, s(w0 , [ M ] =
w0 . Since u X (w0 ) ≥ u X (w0 ), w0 ∈ ϕ[ M ].
Reflexivity entails that some formulas are satisfied only by infinite models.
P ROPOSITION 4.3. There is ϕ ∈ L(P, S) such that ϕ is satisfied by some infinite reflexive
model but by no finite reflexive model.
Proof. Let ϕ be the conjunction of the following formulas.
(4.4) (a) ( p → ( p ≺ X ¬ p))
(b) (¬ p → (¬ p ≺ X p))
It is easy to verify that ϕ is satisfied by a model whose worlds form an ω-sequence
when ordered by u X , and which alternate between satisfying p and ¬ p. On the other hand,
suppose for a contradiction that ϕ is satisfied by finite model M = (W, s, u, t). Then
some w0 ∈ W has maximum u X utility. Suppose that w0 satisfies p (the other case is
parallel). Then (4.4)a and Reflexivity imply that there is w1 ∈ W satisfying ¬ p such that
u X (w0 ) < u X (w1 ). This contradicts the choice of w0 as having maximum u X utility. 4.2. Regularity. If you think that living in Boston is most similar to your current
situation among the set of all addresses in New England then shouldn’t you think that
living in Boston is most similar to your current situation among the set of all addresses
in Massachusetts? A similar principle is standardly applied to choice (Sen, 1971) even
though its violation has been documented in several empirical studies (e.g., Payne & Puto,
1982; Tentori et al., 2001). In the present setting, we are led to the following constraint on
selection.
D EFINITION 4.5. M is regular just in case for all w ∈ W, nonempty A ⊆ B ⊆ W, and
w1 ∈ A: If s(w, B) = w1 then s(w, A) = w1 .
Regularity validates the formula appearing in the next proposition. An instance is this: If
buying either a Ford or a Chevy makes more sense than buying a Toyota then either it
makes more sense to buy a Ford than a Toyota, or it makes more sense to buy a Chevy than
a Toyota (or both).
P ROPOSITION 4.6. Let ϕ be (( p ∨ q) X r ) → (( p X r ) ∨ (q X r )). Then ϕ is invalid
but valid in the class of regular models.
Proof. A countermodel for ϕ is easy to devise. To show validity in the regular models,
suppose that M is regular, and let w ∈ (( p ∨ q) X r )[ M ] be given. Then there are
w1 , w2 ∈ W with:
(4.7) (a) w1 = s(w, ( p ∨ q)[ M ]),
(b) w2 = s(w, r [ M ]), and
(c) u X (w1 ) > u X (w2 ).
PREFERENCE BASED ON REASONS
129
By (4.7)a, either w1 ∈ t ( p) or w1 ∈ t (q), say the latter (the other case is parallel). Since
q[ M ] ⊆ ( p ∨ q)[ M ], it follows from regularity that w1 = s(w, q[ M ]). In view of
(4.7)b and c, w ∈ (q X r )[ M ].
The combination of reflexivity and regularity validates the following formula, which
exhibits modal embedding.
(( p ≺1 ) 2 (q ≺1 )) → (¬ p 2 ¬q).
(3.7)
For an instance, suppose that p, q represent plans for new shopping malls, and that u 1 , u 2
measure their political and ecological interest, respectively. Then (3.7) asserts: If it is ecologically better for p than for q to politically backfire then abstaining from p is ecologically
better than abstaining from q.
P ROPOSITION 4.9. Formula (3.7) is invalid but valid in the class of models that are
reflexive and regular.
Proof. The invalidity of (3.7) is easy to verify. For validity in the class of reflexive regular
models, let reflexive regular model M = (W, s, u, t) and w ∈ W be given. Suppose that:
(4.10) w ∈ (( p ≺1 ) 2 (q ≺1 ))[ M ].
We must show:
(4.11) w ∈ (¬ p 2 ¬q)[ M ].
By (4.10), there are w1 , w2 ∈ W with:
(4.12) (a) 1 )M) = ∅,
(b) w1 = s(w, ( p ≺1 )[ M ]),
(c) w2 = s(w, (q ≺1 )[ M ]),
(d) u 2 (w1 ) > u 2 (w2 ).
By reflexivity, it is easy to verify:
(4.13) (a) ( p ≺1 )[ M ] ⊆ ¬ p[ M ],
(b) (q ≺1 )[ M ] ⊆ ¬q[ M ].
So by (4.12)a,b and (4.13), we have ¬ p[ M ] = ∅ and ¬q[ M ] = ∅. Hence there are
w1∗ , w2∗ ∈ W with:
(4.14) (a) w1∗ = s(w, ¬ p[ M ]),
(b) w2∗ = s(w, ¬q[ M ]).
But by (4.12)a, (4.13)a, (4.14)a, and regularity, w1∗ = w1 . Likewise, by (4.12)b, (4.13)b,
(4.14)b, and regularity, w2∗ = w2 . Thus, (4.12)c implies u 2 (w1∗ ) > u 2 (w2∗ ) which together
with (4.14) yields (4.11).
An alternative formulation of regularity is given by the following definition and proposition.
D EFINITION 4.15. M is locally lexicographic just in case for every v ∈ W there is a well
order Rv of W such that for all A ⊆ W, s(v, A) is the Rv -least member of A.
P ROPOSITION 4.16. A model is regular if and only if it is locally lexicographic.
130
DANIEL OSHERSON AND SCOTT WEINSTEIN
Proof. The right-to-left direction is immediate. For the other direction we proceed as
follows. Suppose (W, s, u, t) is regular and v ∈ W. Define by transfinite recursion a wellordering Rv of W as follows. For every ordinal α let wα = s(v, W − {wβ |β < α}).
[So w0 = s(v, W).] Now let s be the following locally lexicographic selector. For every
nonempty proposition A, s (v, A) = wα where α is the least ordinal γ such that wγ ∈ A.
It suffices to show that for every nonempty proposition A, s(v, A) = s (v, A). Let
s (v, A) = wα . By our construction, A ⊆ (W − {wβ |β < α}). By definition, s(v, W −
{wβ |β < α}) = wα . But wα ∈ A, so by the regularity of s, s(v, A) = wα .
4.3. Metric selection. A natural way to express the thought that selection chooses
“nearby” worlds is via a metric on W. The following definition gives form to this idea.
D EFINITION 4.17. A selection function s over W is metrizable just in case there is a metric
d : W × W → such that for all w ∈ W and ∅ = A ⊆ W, s(w, A) is the unique d-closest
member of A to w. M = (W, s, u, t) is metric just in case s is metrizable.
Note that s is metrizable only if d-closest worlds exist (there are no chains of worlds
ever d-closer to w).
E XAMPLE 4.18. The following model is reflexive and regular but not metric. Let W =
{w0 , w1 , w2 }, and let s be the (unique) reflexive selection function that satisfies:
s(w0 , {w1 , w2 })
s(w1 , {w0 , w2 })
= w1
= w2
s(w2 , {w0 , w1 })
= w0
It is easy to see that s must also be regular. But s is not metrizable since otherwise the
three selections imply d(w0 , w1 ) < d(w0 , w2 ), d(w1 , w2 ) < d(w0 , w1 ), and d(w0 , w2 ) <
d(w1 , w2 ), for some distance metric d. But these inequalities yield d(w0 , w2 ) < d(w0 , w2 ),
contradiction.
As a straightforward consequence of Definition 4.17, we have:
P ROPOSITION 4.19. Every metric model is reflexive and regular. Also, every metric model
is countable.
More consequentially, the next theorem shows that notwithstanding the fact that the metric
models are a proper subset of the reflexive and regular models, metric validity reduces to
reflexive regular validity. See the Appendix for proof.
T HEOREM 4.20. Every formula which is valid in the class of metric models, is valid in the
class of reflexive and regular models.
4.4. Lexicographic ordering. Let us consider another way to strengthen regularity.
D EFINITION 4.21. M is lexicographic just in case there is a well order R of W such that
for all w ∈ W and A ⊆ W, s(w, A) is the R-least member of A.
Thus, the definition provides a uniform version of the local lexicographic property formulated in Definition 4.15. All lexicographic models are regular but not vice versa. The
added constraint imposed by the lexicographic property validates some additional formulas
as shown by the following proposition (whose proof is elementary).
PREFERENCE BASED ON REASONS
131
P ROPOSITION 4.22. Let ϕ be ( p X q) → ( p X q). Then ϕ is false in some regular
model but valid in the class of lexicographic models.
A natural generalization of lexicographic ordering may be defined as follows.
D EFINITION 4.23. A selection function s over W is proposition driven just in case for all
w1 , w2 ∈ W and ∅ = A ⊆ W, s(w1 , A) = s(w2 , A).
That is, proposition driven selection functions ignore their first arguments. Lexicographic
ordering implies proposition drivenness; the next proposition shows the former to be a
stronger condition than the latter.
P ROPOSITION 4.24. There is a formula satisfiable in the class of proposition driven models
but not in the class of lexicographic models.
Proof. Let ϕ be the conjunction of the following formulas.
( p ∨ q) X p ≺X q ≺ X .
It is easy to verify that no regular model satisfies ϕ but that some proposition driven
model does. Since lexicographic ordering implies regularity, the proposition follows immediately.
The next proposition is a corollary to Proposition 4.16.
P ROPOSITION 4.25. A model is proposition driven and regular if and only if it is lexicographic.
§5. Stronger theories that are not based on frames. We now consider properties
of models that cannot be defined just in terms of (W, s), the frame of a model. The
background signature (P, S) may thus be expected to interact with the validity of formulas
in classes of models specified by these properties.
5.1. Proximity. Intuitively, a selection function applied to a world w and nonempty
proposition A should pick a member w1 of A that is “near” or “similar” to w. One way
to articulate this idea is to require that the two worlds differ minimally in the sets of
propositional variables that each makes true. The following notation helps us formulate
this idea. For w ∈ W, let t −1 (w) = {p ∈ P | w ∈ t ( p)}. That is, t −1 (w) is the set
of propositional variables that M satisfies at w. For sets S, T , let S T denote their
symmetric difference (S \ T ) ∪ (T \ S). Then the idea of selecting “nearby worlds” can be
rendered as follows.
D EFINITION 5.1. M is proximal just in case the following condition is met, for all w ∈ W
and all nonempty propositions A ⊆ W.
If s(w, A) = w1 then there is no w2 ∈ A such that t −1 (w) t −1 (w2 ) ⊂ t −1 (w) t −1 (w1 ).
For example, suppose that t −1 (w) = {p, q}, t −1 (w1 ) = {p, r }, and t −1 (w2 ) = {p, q, r }.
Let A = {w1 , w2 }. Then s violates proximity if s(w, A) = w1 since t −1 (w) t −1 (w2 ) =
{r } ⊂ {q, r } = t −1 (w) t −1 (w1 ).
For the next two propositions, we rely on an hypothesis about our signature, namely, that
P = {p, q, r }. In conjunction with regularity, proximity validates a formula reminiscent of
the sure thing principle (Savage, 1954).
132
DANIEL OSHERSON AND SCOTT WEINSTEIN
P ROPOSITION 5.2. Let ϕ be
((( p ∧ r ) X (q ∧ r )) ∧ (( p ∧ ¬r ) X (q ∧ ¬r ))) → ( p X q).
Then ϕ is invalid in the class of regular and in the class of proximal models but valid in the
class of models that are both regular and proximal.
An instance of ϕ is the following. If one has better reason to vacation in Florence during an
Italian transport strike than to vacation in Rome during such a transport strike, and if one
has better reason to vacation in Florence with no transport strike than to vacation in Rome
with no such strike then one has better reason to vacation in Florence than in Rome.
Proof of Proposition 5.2. Construction of the needed countermodels is left for the reader.
Suppose that M is regular and proximal with w ∈ W. Either w ∈ t (r ) or w ∈ t (r ); assume
the former (the argument is parallel in the other case). There is nothing left to prove unless
the following statements are true [since otherwise the left conjunct in the antecedent of ϕ
is false; see Definition 3.5d].
(5.3) (a) t ( p) ∩ t (r ) = ∅
(b) t (q) ∩ t (r ) = ∅.
By (5.3), p[ M ] = ∅ and q[ M ] = ∅. So let w1 , w2 ∈ W be such that:
(5.4) (a) w1 = s(w, p[ M ])
(b) w2 = s(w, q[ M ]).
Since w ∈ t (r ), (5.3)a, (5.4)a, and proximity imply w1 ∈ t ( p) ∩ t (r ). Hence, w1 ∈
( p ∧ r )[ M ] ⊆ p[ M ], so regularity implies w1 = s(w, ( p ∧ r )[ M ]). Likewise, w2 =
s(w, (q ∧ r )[ M ]). From ( p ∧ r ) X (q ∧ r ) we infer u X (w1 ) > u X (w2 ) which in view
of (5.4) implies p X q. Thus w ∈ ϕ[ M ].
Similar reasoning suffices to prove:
P ROPOSITION 5.5. Let ϕ be
( p ∧ (( p ∧ q) X r )) → (q X r ).
Then ϕ is invalid in the class of regular and in the class of proximal models but valid in the
class of models that are both regular and proximal.
For an instance of this formula, suppose that you have a greater gustatory interest in
ham and eggs than just oatmeal. Then if you already have ham, you’ll be more interested
in eggs than just oatmeal.
5.2. Extensionality, saturation, and perfection. We next consider the relation between
worlds and the propositional variables they satisfy. The following condition requires that
distinct worlds don’t make the same variables true.
D EFINITION 5.6. M is extensional just in case for all w1 , w2 ∈ W, {v ∈ P | w1 ∈ t (v)} =
{v ∈ P | w2 ∈ t (v)} implies w1 = w2 .
Observe that if P is finite then every extensional model is finite. Also, it is easy to see that
every proximal extensional model is reflexive. Hence, in a finite signature, no such model
satisfies the conjunction of Formulas (4.4)(a) and (b). If P = {p} then obviously the invalid
formula ( p ≈ X ¬ p) → ( p ≈ X ) is valid in the extensional models.
If every subset of variables inhabits some world, the model may be called “saturated.”
PREFERENCE BASED ON REASONS
133
D EFINITION 5.7. M is saturated just in case for all T ⊆ P there is w ∈ W with {v ∈ P |
w ∈ t (v)} = T .
D EFINITION 5.8. M is perfect just in case M is both extensional and saturated.
In a perfect model, W can be identified with the power set of P. The combination of
perfection and proximity has consequences for the “contraposition” of reasons, as in ( p X
q) → (¬q X ¬ p). This formula is plausible at first sight; for it seems that if p is u X superior to q then u X also favors q rather than p failing to hold. Thus, keeping a promise
is morally superior to teasing the infirm hence not teasing the infirm should be morally
superior to not keeping a promise, which it is. Closer inspection, however, reveals that
only a weaker form of contraposition can be maintained.
P ROPOSITION 5.9. Let C be the class of perfect and proximal models. Then ( p X q) →
(¬q X ¬ p) is not valid in C. However, ((¬ p ∧ ¬q) → ( p X q)) is valid in a given
model of C iff (( p ∧ q) → (¬q X ¬ p)) is valid in the same model.
Proof. We demonstrate the left-to-right direction in the second part of the proposition.
Let M ∈ C be given, and suppose that:
(5.10) (¬ p ∧ ¬q) → ( p X q) is valid in M.
By saturation, let w ∈ t ( p) ∩ t (q). By saturation again, there are w1 , w2 ∈ W with:
(5.11) (a) w1 = s(w, ¬q[ M ]), and
(b) w2 = s(w, ¬ p[ M ]).
To complete the proof it suffices to show that:
(5.12) u X (w1 ) > u X (w2 ).
By (5.11), proximity, and perfection:
(5.13) (a) w1 satisfies the same variables as w, except for q.
(b) w2 satisfies the same variables as w, except for p.
By perfection, there is w∗ ∈ W that satisfies the same subset of P as w except for p, q.
That is, w ∗ falsifies p and q but otherwise agrees with w. Hence by (5.10), w ∗ ∈ ( p X
q)[ M ]. So there are w1 , w2 ∈ W with:
(5.14) (a) w1 = s(w ∗ , p[ M ]),
(b) w2 = s(w ∗ , q[ M ]), and
(c) u X (w1 ) > u X (w2 ).
By (5.14)a,b, proximity, and perfection:
(5.15) (a) w1 satisfies the same variables as w∗ , except for p.
(b) w2 satisfies the same variables as w∗ , except for q.
From (5.13), (5.15), and perfection, w1 = w1 and w2 = w2 . Therefore, (5.12) follows
from (5.14)c.
134
DANIEL OSHERSON AND SCOTT WEINSTEIN
5.3. Conditions on the utility function. We now consider different ways that utilities
can be combined. This topic is at the heart of the relation between reasons and preference.
For as noted earlier, we conceive preference for ϕ over ψ to be represented
by ϕ S ψ,
that is, taking account of all reasons in play. (Here it is assumed that S ∈ S.) We start
with the most basic condition on utility aggregation, namely, that u X depends on just the
u i indexed by X .
D EFINITION 5.16. Let finite X ∈ S be given. Model M is local for X just in case:
(a) for all i ∈ X , {i} ∈ S and
(b) there is a function g from finite subsets of to such that for all w ∈ W, u X (w) =
g(u i (w) | i ∈ X ).
In this case, we call ϕ ∈ L(P, S) g-valid if ϕ is true in the class of models for which u X is
computed via g.
For example, locality prevents u {1,2} (w) from depending on u 3 (w). It is easy to see
that the following formula is valid in the class of {1, 2}-local models but false in some
non{1, 2}-local model.
(( p ≈1 p ) ∧ (q ≈1 q ) ∧ ( p ≈2 p ) ∧ (q ≈2 q )) → (( p ≈{1,2} q) ↔ ( p ≈{1,2} q )).
Candidates for g in Definition 5.16 include:
u X (w) =
average{u i (w) | i ∈ X }
median{u i (w) | i ∈ X }
minimum{u i (w) | i ∈ X }
maximum{u i (w) | i ∈ X }
Formulas separate some of these locality classes. For example, the following schema is
average-valid but neither min- nor max-valid with respect to {i, j}.
((ϕ i ψ) ∧ (ϕ ≈ j ψ)) → (ϕ {i, j} ψ).
To see that the schema is not min-valid, take u j to assign identical numbers to all worlds,
much smaller than the numbers that u i assigns. Do the reverse for a countermodel to maxvalidity.
Next is a schema that is min-valid and max-valid but not average-valid.
(ϕ ≈{i, j,k} ψ) → ((ϕ ≈{i, j} ψ) ∨ (ϕ ≈{i,k} ψ) ∨ (ϕ ≈{j,k} ψ)).
For a countermodel to the formula with respect to averaging, let
w1 , w2 be the worlds attained through ϕ, ψ, respectively, and let the
i, j, k utilities be given in the accompanying table.
Observe that utility aggregation has so far been monotonic in the
following sense.
D EFINITION 5.17. Let X = {x1 . . . xn } ∈ S be given, where also
{x1 }, . . . , {xn } ∈ S. A model M is monotone for X just in case for
all ϕ, ψ,
((ϕ x1 ψ) ∧ · · · ∧ (ϕ xn ψ)) → (ϕ X ψ)
is valid in M.
i
j
k
w1
2
2
2
w2
0
3
3
PREFERENCE BASED ON REASONS
135
The four functions discussed above are consistent with monotonicity but it is easy to
imagine circumstances in which nonmonotonic aggregation takes place. For example, you
might prefer to spend time with people of luxuriant life style (they’re more fun), encoded
in u 1 , and also prefer people who espouse asceticism and self-restraint (they’re more
admirable), encoded in u 2 . The two utility functions considered individually might order
Jim above Jack as dinner partners but u 1,2 will reverse the preference if it is sensitive to
Jim’s hypocrisy.
§6. Decidability and compactness. The present section offers three theorems about
the compactness and decidability of satisfiability (hence, about the decidability of validity
as well). For this purpose, we fix a signature (P, S) in which P is an initial segment of N,
and S is a set of finite subsets of N. The first theorem concerns satisfiability with respect to
the class of all models.
T HEOREM 6.1. The set of satisfiable formulas of L(P, S) is decidable.
Adjustments to the proof of Theorem 6.1 verify the following corollaries.
C OROLLARY 6.2. If a formula of L(P, S) is satisfiable then it is satisfied in a finite model
(i.e., in a model with finitely many worlds).
C OROLLARY 6.3. The set of formulas of L(P, S) that are satisfiable in the class of reflexive
models is decidable.
Corollary 6.2 may be contrasted with Proposition 4.3, stating that some formulas can be
satisfied by a reflexive model only if the model contains infinitely many worlds.
The second theorem bears on lexicographic ordering in the sense of Definition 4.21, and
on proposition drivenness in the sense of Definition 4.23.
T HEOREM 6.4. The set of formulas of L(P, S) that are satisfiable in the class of lexicographic models is decidable, as is the set of formulas that are satisfiable in the class of
proposition driven models. Indeed, both sets of formulas are NP-complete.
The final theorem affirms that satisfiability with respect to the class of all models is
countably compact. We call a collection ⊆ L of formulas “satisfiable” just in case there
is a model M that satisfies every member of at a common point, that is, just in case:
{ϕ[ M ] | ϕ ∈ } = ∅.
T HEOREM 6.5. Suppose that signature (P, S) is countable, and let ⊆ L(P, S) be given.
Then is satisfiable if and only if every finite subset of is satisfiable.
On the other hand, if either P or S is uncountable then compactness breaks down. Proofs
of the theorems are provided in the Appendix.
§7. Generalized frames. In our theory, ϕ X ψ can be understood as asserting that
u X assigns at least as much value to the proposition expressed by ϕ as to the proposition
expressed by ψ. The latter two propositions are represented by elements of each, picked
out as a function of the world at which the formula is evaluated. A natural generalization
is to compare the value of propositions directly, without recourse to individual worlds as
representatives. We explore this idea in the present section. Let (P, S) be our background
136
DANIEL OSHERSON AND SCOTT WEINSTEIN
signature, and recall that a total preorder is transitive, connected, and reflexive over its
domain.
D EFINITION 7.1. Let a set W of worlds be given.
(a) By a value-ordering for W and S is meant a function v from W × S to the set of
total preorders over the class of nonempty subsets of W. We call the pair (W, v) a
generalized frame.
(b) Let a truth-assignment t and a value-ordering v for W and S be given. Then (W, t, v)
is a generalized model.
Intuitively, a value-ordering arranges propositions by utility, relative to index X ∈ S and
vantage point w ∈ W. The semantics of generalized models is given by Definition 3.5
with the following substitution for Clause 3.5d. Let ϕ ∈ L(P, S) and generalized model
M = (W, t, v) for (P, S) be given.
3.5d If ϕ has the form (θ X ψ) for X ∈ S, then ϕ[ M ] = ∅ if either
θ[ M ] = ∅ or ψ[ M ] = ∅. Otherwise:
ϕ[ M ] = {w ∈ W | θ [ M ] comes no earlier than ψ[ M ] in v(w, X )}.
Now let model M = (W, s, u, t) be given. Then a value-ordering v is induced by the
following condition. For w ∈ W, X ∈ S, and nonempty A, B ⊆ W, A is (weakly) ordered
after B iff u X (w A ) ≥ u X (w B ) where w A = s(w, A) and w B = s(w, B). (The truthassignment t plays no role.) We call (W, v) the generalized frame induced by M.
Given a model (W, s, u, t), w ∈ W, and nonempty A ⊆ W, there is w0 ∈ W with
u X (s(w, A)) = u X (s(w, {w0 })), namely, w0 = s(w, A). So we have:
L EMMA 7.2. Let value-ordering v be induced by model (W, s, u, t). Then for all w ∈ W
and X ∈ S, every equivalence class in v(w, X ) contains a singleton set.
We have the following immediate consequence, which shows that some generalized frames
cannot be induced by models.
P ROPOSITION 7.3. Let W contain at least two worlds. Let value-ordering v be such that
for some w ∈ W and X ∈ S, either
(a) v(w, X ) refines ⊂ over the field of nonempty subsets of W or
(b) v(w, X ) is a strict linear order over the nonempty subsets of W.
Then (W, v) is not induced by any model.
On the other hand, the next proposition shows that some interesting classes of generalized
frames can be characterized in L(P, S). We say that ϕ ∈ L(P, S) is valid in a generalized
frame (W, v) in case ϕ is true in every generalized model of form (W, t, v).
P ROPOSITION 7.4. Let X ∈ S be given. There are ϕ1 , ϕ2 ∈ L(P, S) such that
(a) ϕ1 is valid in a generalized frame (W, v) if and only if for all w ∈ W, v(w, X ) refines
⊂ over the nonempty subsets of W;
(b) ϕ2 is valid in a generalized frame (W, v) if and only if for all w ∈ W, v(w, X ) is a
strict linear order over the nonempty subsets of W.
Proof. It is easy to verify the proposition with the following choices of ϕ1 , ϕ2 , respectively.
PREFERENCE BASED ON REASONS
137
(( p → q) ∧ ¬(q → p) ∧ ♦ p) → p ≺ X q
¬( p ↔ q) → (( p ≺ X q) ∨ (q ≺ X p))
In Osherson & Weinstein (To appear), we provide an axiomatization of the set of formulas that are valid in generalized frames and establish that a formula is valid in the class of
generalized frames if and only if it is valid in the class of frames.
§8. Discussion. The foregoing investigation raises many questions and avenues for
further research. We indicate some directions.
8.1. Utility. Suppose that distinct {i}, {j}, {k} ∈ S. For all ϕ, θ ∈ L(P, S), let:
def
(ϕ V θ) = ((ϕ i θ) ∧ (ϕ j θ)) ∨ ((ϕ i θ) ∧ (ϕ k θ)) ∨ ((ϕ j θ) ∧ (ϕ k θ)).
Then ϕ V θ is true if a majority of the utility scales i, j, k are favorable to ϕ compared to
θ . Observe that ((ϕ V θ) ∧ (θ V ψ)) → (ϕ V ψ) (transitivity) is not guaranteed in a given
model inasmuch as the utility scales u i , u j , u k might embody a voting cycle (see Johnson,
1998). Therefore, V cannot itself be represented by a utility scale. The following matter
thus merits exploration.
O PEN Q UESTION 8.1. Suppose that {i, j, k} ∈ S. Under what conditions does ϕ V θ imply
ϕ i, j,k θ, and vice versa?
The voting operator V might best be analyzed in the context of a generalization of our
approach to utility. Instead of utility scales corresponding to each X ∈ S, we may posit
relations R X ⊆ W × W. In this setup, θ X ψ is true at w ∈ W just in case (s(w, θ [ M ]),
s(w, ψ[ M ])) ∈ R X . Such relations R X could vary in their order-theoretic properties (e.g.,
transitivity) as well as in their connection to relations Ri with i ∈ X . This perspective
might allow the remarkable results developed in Andréka et al. (2002), about combining preference relations, to shed light on the logic of reasons. In Osherson & Weinstein
(Forthcoming), we axiomatize the sets of valid formulas that arise under various choices
of relation R.
Questions also remain about the classes of utility functions defined in Section 5.3. Can
any of them be uniquely characterized by a set of formulas? Even the less ambitious
problem of separating utility functions is currently unresolved. For example, the following
question was left open.
O PEN Q UESTION 8.2. Is there ϕ ∈ L(P, S) that is minimum-valid but not maximum-valid
(and vice versa)?
8.2. Selection. Additional conditions on selection functions remain to be investigated.
For example, an alternative concept of selection allows more than one world to be “nearest”
to a target. To express this idea, we replace Definition 3.1 with the following.
D EFINITION 8.3. A wide selection function s over W is a mapping from W × {A ⊆ W |
A = ∅} to the power set of W such that for all w ∈ W and ∅ = A ⊆ W, ∅ = s(w, A) ⊆ A.
Selection functions in the original sense of Definition 3.1 can now be seen as the special
case in which only singleton sets are returned. To satisfy a formula θ X ψ in the context
138
DANIEL OSHERSON AND SCOTT WEINSTEIN
of a wide selection function, we may require that some nearby θ-world is weakly X -better
than some nearby ψ-world, or that all of them are, etc. The consequences of these options
have yet to be explored.
8.3. Updating. Suppose you live in a model M = (W, s, u, t) but wish to take on
board ϕ ∈ L as an assumption. We take this to mean that ϕ will be made true in all
worlds of some successor model M = (W , s , u , t ) that is the natural ϕ-update to M.
(Updating is analyzed from a graph-theoretic perspective in Andréka et al., 2002; van
Benthem & Liu, 2007.)
If ϕ is boolean, updating M seems easy: set W = {w ∈ W | w | ϕ}, and let s ,
u , t be the obvious reducts of s, u, and t to W . (Updating in this sense is not defined if
W = ∅.) But if ϕ has a modal connective, matters are not straightforward. Consider the
following choice for M, where (P, S) = ({p, q}, {{i}}).
W = {w1 , w2 , w3 , w4 }
t ( p) = {w2 , w4 } t (q) = {w3 }
u i (w4 ) < u i (w3 ) < u i (w2 ) < u i (w1 )
s(w1 , p[ M ]) = w2
s(w1 , q[ M ]) = w3
s(w2 , p[ M ]) = s(w3 , p[ M ]) = s(w4 , p[ M ]) = w4
s(w2 , q[ M ]) = s(w3 , q[ M ]) = s(w4 , q[ M ]) = w3
For ϕ := p ≺i q to be true throughout M , it suffices to remove w1 from W. But
since s(w1 , {w4 }) must equal w4 , it is easy to verify that removing w2 from W also
suffices for the same purpose. Updating in the general case thus requires choice among
successor models, in a sense familiar from the theory of belief revision (Gärdenfors, 1988).
Investigation of the matter might usefully address the following issue. Given a proposed
updating operator ‡ and a class C of models with (say) the regularity property, for which
ϕ ∈ L (if any) is {M ‡ ϕ | M ∈ C} guaranteed to be regular?
§9. Appendix: Proof of Theorem 4.20. The theorem is an immediate corollary to
the following lemma, the proof of which involves the notion of modal depth, defined as
follows.
D EFINITION 9.1. We define μ(ϕ), the modal depth of ϕ, by recursion on ϕ ∈ L(P, S) as
follows.
⎧
0
if ϕ ∈ P
⎪
⎪
⎨
μ(ψ)
if ϕ = ¬ψ
μ(ϕ) =
max{μ(ψ), μ(θ)}
if ϕ = (ψ ∧ θ)
⎪
⎪
⎩
max{μ(ψ), μ(θ)} + 1 if ϕ = (ψ X θ)
L EMMA 9.2. Suppose P and S are finite. For every ref and regular model M = (W, s, u, t),
w ∈ W, and n ∈ N, there is a metric model M∗ = (W∗ , s ∗ , u ∗ , t ∗ ) and a world w ∗ ∈ W∗
such that for every ϕ ∈ L(P, S), if μ(ϕ) ≤ n, then
w ∈ ϕ[ M ] if and only if w ∗ ∈ ϕ[ M∗ ].
PREFERENCE BASED ON REASONS
139
Proof. Let ref and regular model M = (W, s, u, t), w ∈ W, and n ∈ N be given. To
prove the proposition, we construct from M a metric model M∗ = (W∗ , s ∗ , u ∗ , t ∗ ) and a
world w ∗ ∈ W∗ such that for every ϕ ∈ L(P, S), if μ(ϕ) ≤ n, then
w ∈ ϕ[ M ] if and only if w∗ ∈ ϕ[ M∗ ].
To reduce notational clutter, we suppress the subscripts on occurrences of and suppress
them likewise on utility functions u. The construction proceeds in three steps. In the first
step, we unravel M at the world w with respect to the modalities generated by the selection
function s and the propositions expressed by formulas of modal depth ≤ n. (Unraveling
is a standard technique in modal logic, see Blackburn et al. 2001, sec. 2.1.) This leads
to a “partial” model M† = (W∗ , s † , u ∗ , t ∗ ) and a world w ∗ ∈ W∗ such that for every
ϕ ∈ L(P, S) with μ(ϕ) ≤ n,
(9.3)
w ∈ ϕ[ M ] if and only if w ∗ ∈ ϕ[ M† ],
in which s † is only defined in its second argument on propositions expressed by formulas
of modal depth < n. In the second step, we define a metric d on W∗ with respect to which
the partial selector s † is metric. Finally, we extend s † to a total selector s ∗ and verify that
d is a metricization of the resulting model M∗ .
S TEP 1: For every k ∈ N, let k = {ϕ[ M ] | μ(ϕ) ≤ k} − {∅}. It follows immediately
from the finiteness of (P, S) that for every k ∈ N, k is finite. For each Z ∈ n−1 , define
a binary relation R Z on W as follows.
For all v, v ∈ W , R Z (v, v ) if and only if s(v, Z ) = v .
We use the “accessibility” relations R Z to define the unraveling of M at w. A finite
sequence w0 , w1 , . . . , wm from W is a world in W∗ if and only if w0 = w and for each
0 ≤ i < m, there is a Z ∈ n−1 such that R Z (wi , wi+1 ). If v ∗ = w0 , w1 , . . . , wm ∈ W∗ ,
we write last(v ∗ ) for wm . For w ∈ W, we use v ∗ w to denote w0 , w1 , . . . , wm , w , the
concatenation of w to the end of v ∗ . Define u ∗ and t ∗ by emulating u and t, namely:
(9.4) For all v ∗ ∈ W∗ ,
(a) u ∗ (v ∗ ) = u(last(v ∗ )) and
(b) v ∗ ∈ t ∗ ( p) if and only if last(v ∗ ) ∈ t ( p), for every p ∈ P.
In order to complete the construction of the partial model M† it remains only to define the
partial selector s † (v ∗ , ψ[ M† ]) for all v ∗ ∈ W∗ and all ψ ∈ L(P, S) of modal depth < n.
We do this by a recursion on j ≤ n which simultaneously defines a sequence of partial
selectors s †j and partial models M†j = (W∗ , s †j , u ∗ , t ∗ ). We begin the recursion by setting
s0† = ∅. Then for 0 ≤ j < n and for ψ of modal depth ≤ j, we let
∗
v s(last(v ∗ ), ψ[ M ]) if s(last(v ∗ ), ψ[ M ]) = last(v ∗ )
†
†
∗
(9.5) s j+1 (v , ψ[ M j ]) =
v∗
otherwise.
In order to justify this inductive definition we must show that for each 0 ≤ j ≤ n,
ψ[ M†j ] is well-defined by Definition 3.5 for each ψ of modal depth ≤ j. For this, it
suffices to show that for each 0 ≤ j < n, s †j+1 is a partial selector, that is, we must show
that
140
DANIEL OSHERSON AND SCOTT WEINSTEIN
(9.6) for every v ∗ ∈ W∗ and ψ, χ ∈ L(P, S) of modal depth ≤ j,
(a) ψ[ M†j ] = χ[ M†j ] if and only if ψ[ M ] = χ[ M ] and
(b) if ψ[ M†j ] = ∅, then s †j+1 (v ∗ , ψ[ M†j ]) ∈ ψ[ M†j ].
Condition (9.6)a guarantees that the definition of s †j+1 via (9.5) is consistent, while
(9.6)b insures that s †j+1 is a selector for propositions expressed by sentences of modal
depth ≤ j. We will establish (9.6) by induction and simultaneously prove by induction that
for all 0 ≤ j ≤ n,
(9.7) for all v ∗ ∈ W∗ and ψ, χ ∈ L(P, S) of modal depth ≤ j,
(a) v ∗ ∈ ψ[ M†j ] if and only if last(v ∗ ) ∈ ψ[ M ], if and only if ψ[ M ] = χ[ M ] and
(b) ψ[ M†j ] = ∅ if and only if ψ[ M ] = ∅.
To begin the induction, let j = 0. Then for all ψ ∈ L(P, S) of modal depth ≤ j, ψ[ M†j ]
is well-defined, since, by Definition 3.5, its value does not depend on s †j . Moreover, in this
case, (9.7)a is a direct consequence of (9.4)b. Next, note that the right-to-left direction
of (9.7)b follows immediately from (9.7)a. For the left-to-right direction, suppose that
ψ[ M ] = ∅. It then follows from the definition of W∗ that there is a v ∗ ∈ W∗ such
that last(v ∗ ) ∈ ψ[ M ]. Therefore, by (9.7)a, ψ[ M†j ] = ∅. Observe next that the right-toleft direction of (9.6)a also follows immediately from (9.7)a. For the left-to-right direction,
suppose that ψ[ M ] = χ [ M ], say, (ψ∧¬χ)[ M ] = ∅; the other case (χ ∧¬ψ)[ M ] = ∅
is handled the same way. Since ψ and χ are sentences of modal depth ≤ j, so is (ψ ∧ ¬χ).
It follows from the definition of W∗ that there is a v ∗ ∈ W∗ such that last(v ∗ ) ∈ ψ[ M ] −
χ[ M ]. Therefore, by (9.7)a, ψ[ M†j ] = χ[ M†j ]. Finally, (9.6)b follows immediately
from (9.5) and (9.7)a.
For the induction step, suppose that (9.6)a,b and (9.7)a,b hold for all v ∗ ∈ W∗ and
ψ, χ ∈ L(P, S) of modal depth ≤ j for some 0 < j < n, and let θ ∈ L(P, S) be a
sentence of modal depth j + 1. It is easy to see from the definition of the models under
construction that ψ[ M†j ] = ψ[ M†j+1 ] for ψ ∈ L(P, S) of modal depth ≤ j. It follows
at once, from the induction hypothesis, that θ [ M†j+1 ] is well-defined by Definition 3.5. In
order to conclude the argument, it suffices to show that (9.7)a holds for θ , since the proofs
of (9.6)a,b and (9.7)b from (9.7)a are exactly parallel to those in the basis of the induction.
Moreover, since θ has modal depth j + 1, θ is a boolean combination of sentences of the
form ψ χ where ψ and χ have modal depth ≤ j. Since it is easy to see that (9.7)a
is preserved under boolean combinations, we may suppose that θ is ψ χ for some ψ
and χ of modal depth ≤ j. In case either ψ[ M†j+1 ] = ∅ or χ[ M†j+1 ] = ∅, it follows
immediately from the induction hypothesis (9.7)b that (9.7)a holds for θ . So suppose that
ψ[ M†j+1 ] = ∅ and χ[ M†j+1 ] = ∅. It then follows immediately from (9.4)a, (9.5), and
the induction hypothesis (9.7)a for ψ and χ that (9.7)a holds for θ .
Finally, let M† = M†n and w∗ = w. It follows at once that (9.3) holds for all ϕ ∈
L(P, S) of modal depth ≤ n.
S TEP 2: We now define a metric d on W∗ which is a metricization of the partial structure
M† , that is,
(9.8) for each ψ ∈ L(P, S) of modal depth < n, and each v ∗ ∈ W∗ , s † (v ∗ , ψ[ M† ]) is
the unique d-nearest world to v ∗ .
PREFERENCE BASED ON REASONS
141
Let † = {ϕ[ M† ] | μ(ϕ) ≤ (n − 1)} − {∅}. For v ∈ W∗ , let Cv = {s † (v, Z ) | Z ∈ † },
and let kv be the cardinality of Cv . Note that W∗ is a finitely branching directed tree T
where v is a child of v if and only if v = vlast(v ). For all v, v ∈ W∗ , v ∈ Cv if and
only if v is a child of v or v = v, hence the out-degree of each world is kv − 1. We first
show that for every v ∈ W∗ there is a strict linear ordering x1v , . . . , xkvv of Cv such that
(9.9) for every Z ∈ † , s † (v, Z ) is the xiv ∈ Z of lowest index.
For the proof of (9.9), observe that the regularity of M implies that for every v ∈ Cv ,
there is a Z ∈ † such that Z ∩ Cv = {v }. Since † is closed under boolean combinations,
it is immediate that for every nonempty Y ⊆ Cv , there is a Z Y ∈ † such that Z Y ∩Cv = Y .
Now, for each v ∈ W∗ , we define the xiv ’s by induction on i ∈ N up to kv . Let x1v = v and
v
= s † (v, Z Cv −{x1v ,...,xiv } ). It follows directly from the definition and
for 1 ≤ i < kv , let xi+1
the reflexivity and regularity of M that (9.9) holds.
As general background, for T an edge-weighted directed tree and v, v vertexes of T ,
we write d(v, v ) for the the ordinary (i.e., weighted path-length) distance between v and
v in the edge-weighted symmetrized tree derived from T . Returning to our tree T , we
write Tm for the restriction of T to its vertexes of height ≤ m and Vm for its set of vertexes
of height exactly m. We proceed to introduce weights on the edges of T via recursion on
Tm . The basis is trivial, since T0 contains no edges. At stage m > 0, for all v ∈ Vm−1
and 1 < i ≤ kv , we choose weights λiv on the edges from v to xiv to satisfy the following
conditions.
v ;
(9.10) (a) if i < kv , then λiv < λi+1
(b) λiv ∈ (1 + 2−(m+1) , 1 + 2−m );
(c) for all v, v , v ∈ Tm , if v = v , then d(v, v ) = d(v, v ).
It is clear that such weights can be chosen. Now (9.8) follows directly from (9.9) and
(9.10)a,b.
S TEP 3: By (9.10)c, we may define a total selector s ∗ extending s † such that for every
v ∈ W∗ and every nonempty X ⊆ W∗ , s ∗ (v, X ) is the unique d-nearest member of X to
v. This completes our definition of the metric structure M∗ . It follows at once from (9.3)
that for all ϕ ∈ L(P, S) of modal depth ≤ n, w ∈ ϕ[ M ] if and only if w∗ ∈ ϕ[ M∗ ]. 2
§10. Appendix: proof of Theorem 6.1. To demonstrate that the set of satisfiable
formulas of L(P, S) is decidable, we apply the well-known “method of mosaics” (see
Blackburn et al. 2001, sec. 6.4). We carry out the construction in some detail.
Let θ ∈ L(P, S) be given. Let be the collection of subformulas of θ , and let Z be
the set of utility indices that appear in θ . We close under one application of negation,
followed by one application of ↔, followed by one application of negation, followed
by one application of 3, followed by one application of negation. The resulting set of
formulas will be called . We say that ⊆ is a Hintikka set (abbreviated H-set) if and
only if
(10.1) (a) for every ¬ϕ ∈ , ϕ ∈ iff ¬ϕ ∈ and
(b) for every (ϕ ∧ ψ) ∈ , (ϕ ∧ ψ) ∈ iff both ϕ ∈ and ψ ∈ .
We let be the collection of all H-sets. Note that if n is the length of θ , then the size
2
c of (and thus of every H-set) is O(n 2 ). Therefore, the size d of is O(2n ). For the
purposes of the next definition, we establish the notational convention that if f is the graph
142
DANIEL OSHERSON AND SCOTT WEINSTEIN
of a partial function, we write f (a) for the b such that a, b ∈ f , when a is in the domain
of f . A brick is a triple , σ, {υ X | X ∈ Z } where
(10.2) (a) is an H-set;
(b) σ is the graph of a partial function from into such that ϕ ∈ σ (ϕ) for every ϕ ∈ on which σ is defined;
(c) for each X ∈ Z , υ X is a function from σ to {i | 1 ≤ i ≤ card(σ )};
(d) if ♦ϕ ∈ , then for some ∈ range(σ ), ϕ ∈ ;
(e) if ♦ϕ ∈ , then for all ∈ range(σ ) ∪ {}, ϕ ∈ ;
(f) (ϕ ↔ ψ) ∈ if and only if σ (ϕ) = σ (ψ);
(g) (ϕ X ψ) ∈ iff υ X (ϕ, σ (ϕ)) ≤ υ X (ψ, σ (ψ)).
Let z be the size of Z . Note that the number b of bricks is O(d c+1 · ccz ).
If β is a brick, we write β1 , β2 , and β3 for the first, second, and third coordinates of β.
A set B of bricks is a mosaic if and only if
(10.3) (a) for all β, β ∈ B, {ϕ | ♦ϕ ∈ β1 } = {ϕ | ♦ϕ ∈ β1 } and
(b) for all β ∈ B and for all ∈ range(β2 ) there is a β ∈ B such that β1 = .
A set B of bricks is a mosaic for θ ∈ L(P, S) if and only if B is a mosaic and for
some β ∈ B, θ ∈ β1 . Note that the number of mosaics is O(2b ) and that it is decidable in
time polynomial in the size of a set B of bricks whether B is a mosaic. It follows that the
decision problem “Does there exist a mosaic for θ ” is in NTIME(b). Theorem 6.1 is thus a
corollary to the following.
P ROPOSITION 10.4. For every θ ∈ L(P, S), θ is satisfiable if and only if there is a mosaic
for θ.
To prove the left to right direction of Proposition 10.4, let satisfiable θ ∈ L(P, S) be
given. For notational convenience we assume that Z = {X }. The generalization to multiple
utility indices is routine. Let model M = (W, s, u, t) satisfy θ and suppose that w0 ∈
θ [ M ]. For each w ∈ W, let w = {ϕ ∈ | w ∈ ϕ[ M ]}. Note that for every w ∈ W,
w is an H-set. Now for each w ∈ W, we construct a brick βw = w , σw , υw , where for
each ϕ ∈ ,
σw (ϕ) = s(w,ϕ[ M ])
and for all ϕ, ψ ∈ w
υw (ϕ, σw (ϕ)) ≤ υw (ψ, σw (ψ)) iff u(s(w, ϕ[ M ])) ≤ u(s(w, ψ[ M ])).
Let B = {βw | w ∈ W}. It is easy to verify that B is a mosaic for θ.
For the right to left direction of Proposition 10.4, suppose that B is a mosaic for θ. We
show how to use the bricks of B to construct a tree-like infinite model M = (W, s, u, t),
the root world w0 of which satisfies θ . The tree underlying the model is both node-labeled
and edge-labeled. We call the node labels “brick-labels” and we call the edge labels
“selector-labels.” The construction of the tree proceeds by induction; at stage i, we construct the worlds of depth i.
At stage 0 we introduce a world w0 and label w0 with some brick β ∈ B such that
θ ∈ β1 (such a brick-label exists, since B is a mosaic for θ).
Let Wn be the set of worlds constructed at stage n. At stage n + 1 we proceed as follows.
For each w ∈ Wn we construct the children of w as follows. Let βw = w , σw , υw be
the brick-label of w. For each ∈ range(σw ) we introduce a child w of w; we label the
PREFERENCE BASED ON REASONS
143
edge from w to w with selector-label σw−1 [](= {ϕ | σw (ϕ) = }) and we choose as
the the brick-label of w some brick β ∈ B with β1 = (by (10.3)ii, such
a brick exists
since B is a mosaic). This completes the definition of Wn+1 . We let W = n Wn . Next we
define u and t.
(a) For every p ∈ P, t ( p) = {w | p ∈ (βw )1 }.
(b) For every w ∈ W u(w ) = υw (), where w and are the unique world and Hintikka
set such that w is a child of w and (βw )1 = .
To complete the definition of the model M we must specify the selection function s.
Recall that is the set of subformulas of θ. For nonempty subsets T of W not of the form
ϕ[M] for some ϕ ∈ , and for each w, we choose s(w, T ) to be an arbitrary element of
T . For ϕ ∈ we wish to define s so that for every w ∈ W, s(w, ϕ[M]) is the unique child
w of w such that ϕ is an element of the selector-label of the edge from w to w , provided
that ϕ[M] is nonempty. Since, in general, ϕ[M] depends on s, we will define s on sets of
the form ϕ[M] by recursion on the logical complexity of ϕ. Simultaneously, we will prove
by induction on logical complexity that for every w ∈ W and ϕ ∈ ,
w ∈ ϕ[M] iff ϕ ∈ (βw )1 ,
thereby completing the proof of the theorem.
Basis: It follows immediately from the definition of t that for every p ∈ and w ∈ W:
(10.5) w ∈ p[M] iff p ∈ (βw )1 .
Now, for each p ∈ and w ∈ W, we let s(w, p[M]) be the unique child w of w such
that p is an element of the selector-label of the edge from w to w , provided that p[M] is
nonempty. In order to secure the legitimacy of this definition we need to verify that
(10.6) (a) s(w, p[M]) ∈ p[M];
(b) p[M] = ∅ iff no selector-label of an edge exiting w contains p;
(c) for all q ∈ , p[M] = q[M] iff p and q are contained in exactly the same selectorlabels of edges exiting w (i.e., either none of them, or a unique one containing both).
Condition (10.6)a follows immediately from (10.5). In order to establish (10.6)b, suppose first that p[M] = ∅. By (10.5), it follows that for all w ∈ W, p ∈ (βw )1 . Hence,
for no child w of w is p ∈ (βw )1 . Therefore, no selector-label of an edge exiting w
contains p. For the converse, suppose that no selector-label of an edge exiting w contains
p. Note that since p ∈ , ¬♦ p ∈ . It follows from (10.2)d that ♦ p ∈ (βw )1 , and thence
from (10.3)a, that for every w ∈ W, ♦ p ∈ (βw )1 . Hence, by (10.1)a, for every w ∈ W,
¬♦ p ∈ (βw )1 . Therefore, by (10.2)e and (10.1)a, for every w ∈ W, p ∈ (βw )1 . Hence,
by (10.5), p[M] = ∅.
In order to establish (10.6)c, suppose first that q ∈ and p[M] = q[M]. By (10.5),
we may conclude that for all w ∈ W, p ∈ (βw )1 iff q ∈ (βw )1 ; the RHS of (10.6)c now
follows immediately from the definition of M (the argument parallels that for the left to
right direction of (10.6)b above). Finally, suppose that p and q are contained in exactly
the same selector-labels of edges exiting w. We may suppose that this set is nonempty,
for otherwise the result follows from (10.6)b. It follows at once that σw ( p) = σw (q).
Note that since p, q ∈ , ( p ↔ q) ∈ ; we may then conclude, by (10.2)6, that
( p ↔ q) ∈ (βw )1 . Hence, by (10.3)a, for all w ∈ W, ( p ↔ q) ∈ (βw )1 . But then,
by (10.1)a,b and (10.2)e, for all w ∈ W, p ∈ (βw )1 iff q ∈ (βw )1 . We may conclude, by
(10.5), that p[M] = q[M].
144
DANIEL OSHERSON AND SCOTT WEINSTEIN
Induction Hypothesis: Suppose that for all w ∈ W,
(10.7) (a) w ∈ ϕ[M] iff ϕ ∈ (βw )1 ;
(b) w ∈ ψ[M] iff ψ ∈ (βw )1 ;
(c) s(w, ϕ[M]) and s(w, ψ[M]) are determined.
Induction Step: It follows immediately from (10.1)a,b and (10.7)a,b that for all w ∈ W,
w ∈ (ϕ ∧ ψ)[M] iff (ϕ ∧ ψ) ∈ (βw )1
and
w ∈ (¬ϕ)[M] iff (¬ϕ) ∈ (βw )1 .
It remains to show that
w ∈ (ϕ ψ)[M] iff (ϕ ψ) ∈ (βw )1 .
Suppose first that w ∈ (ϕ ψ)[M]. Then, s(w, ϕ[M]) and s(w, ψ[M]) are both
defined and υw (s(w, ϕ[M])) = u(s(w, ϕ[M])) ≤ u(s(w, ψ[M])) = υw (s(w, ψ[M])).
It follows at once, by (10.2)g and (10.7)a,b, that (ϕ ψ) ∈ (βw )1 . Suppose, on the other
hand, that w ∈ (ϕ ψ)[M]. Then either at least one of ϕ[M] or ψ[M] is empty, or
υw (s(w, ψ[M])) = u(s(w, ψ[M])) < u(s(w, ϕ[M])) = υw (s(w, ϕ[M])). In either
case, it follows from (10.7)a,b, (10.1)a, and (10.2)g, that (ϕ ψ) ∈ (βw )1 .
The extension of the definition of s to (ϕ ∧ ψ)[M], (¬ϕ)[M], and (ϕ ψ)[M] is
justified exactly as in the basis of the induction.
2
The above argument may be adapted to establish Corollaries 6.2 and 6.3.
Proof of Corollary 6.2. We modify the construction of (W, s, u, t) in the argument for
the right to left direction of
Proposition 10.4 to build a finite model satisfying θ from a
mosaic B for θ . Let W = n Wn be the set of worlds constructed in the proof above, and
let n be the first stage such that for every w ∈ W there is an m < n and a w ∈ Wm such
that βw = βw . We now close the construction of (W,
s, u, t) at stage n + 1 by choosing
the children of each w ∈ Wn to be suitable worlds in m≤n Wm that satisfy the conditions
in the construction of (W, s, u, t).
Proof of Corollary 6.3. We modify the definition of brick as follows. A reflexive brick
B = , σ, υ is a brick that satisfies the following additional condition:
(10.8) for all ϕ ∈ , s(ϕ) = .
A reflexive mosaic is a mosaic composed of reflexive bricks; a reflexive mosaic for θ is
defined similarly. Corollary 6.3 follows from the next proposition.
P ROPOSITION 10.9. For every θ ∈ L(P, S), θ is satisfiable in a reflexive model if and only
if there is a reflexive mosaic for θ .
The proof of Proposition 10.9 is a straightforward adaptation of the proof of Proposition
10.4. The only subtlety is that in the definition of the tree-like model (W, s, u, t) we can
no longer define u in advance, but must define it by recursion following the recursive
construction of the tree. We extend the definition of u to the children of a world w by
choosing rational values for the children in such a way as to establish an isomorphism with
the order induced by υw , remembering that w will be chosen as a child of w in the obvious
way, so that u must retain the value for w that was determined at stage n.
PREFERENCE BASED ON REASONS
145
§11. Appendix: proof of Theorem 6.4. Both sets of formulas are NP-hard since the
satisfiability problem for nonmodal sentential logic is ptime-reducible to each. A straightforward application of the mosaic method yields the conclusion that each is in NP.
2
§12. Appendix: proof of Theorem 6.5. We derive the compactness theorem for
L(P, S), where the signature (P, S) is countable, as a corollary to the compactness theorem
for first-order logic. Our argument follows a standard strategy which proceeds via translating a modal language into a (fragment of) first-order logic. The translation essentially codifies the definition of satisfaction over some class of relational frames. In our case, this direct
strategy requires modification since our “frames” are not (first-order) relational structures,
in particular, the selection function has a “type 1” argument, the proposition from which
a salient confirming representative is chosen. Moreover, the utility functions, whose range
is the set of real numbers, present another obstacle to smooth “first-orderization” in the
context of our compactness argument. To overcome these difficulties, the first step in our
compactness proof is to “compile” a structure M = (W, s, u, t) into a relational structure
FM and to translate each sentence ϕ ∈ L(P, S) to a first-order formula ϕ † (x) with one
free variable such that for all w ∈ W,
(12.1) w ∈ ϕ[M] iff FM | ϕ † [w].
Given M = (W, s, u, t) of signature (P, S), we define FM . The signature of FM
consists of a unary relation symbol Q p , for each p ∈ P; a binary relation symbol ≤ X , for
each X ∈ S; and a binary relation symbol Rϕ , for each sentence ϕ of L(P, S).
(12.2) The interpretation of each relation symbol in the signature of FM is defined as
follows (we suppress the superscript FM on each relation symbol):
(a) Q p = t ( p);
(b) for all w, w ∈ W, w ≤ X w iff u X (w) ≤ u x (w );
(c) for all w, w ∈ W, Rϕ (w, w ) iff w = s(w, ϕ[M]).
Note that (12.2)c implies that Rϕ is the empty relation if and only if ϕ[M] = ∅.
(12.3) We now define, by recursion, for each sentence ϕ of L(P, S), its translation ϕ † (x),
a formula with one free variable in the first-order language of FM .
(a) p † = Q p (x);
(b) (ϕ ∧ ψ)† = ϕ † (x) ∧ ψ † (x);
(c) (¬ϕ)† = ¬ϕ † (x);
(d) (ϕ X ψ)† = (∃y)(∃z)(ϕ † (y) ∧ ψ † (z) ∧ Rϕ (x, y) ∧ Rψ (x, z) ∧ y ≤ X z).
On the basis of (12.2) and (12.3), it is now easy to verify (12.1).
Next, we describe a first-order theory D in the signature of FM such that
(12.4) for every M, FM | D
and
(12.5) for every countable first-order structure A, if A | D, then for some M, A = FM .
We proceed to describe D.
146
DANIEL OSHERSON AND SCOTT WEINSTEIN
(12.6) D consists of the following first-order sentences.
(a)
(b)
(c)
(d)
(e)
(f)
(∀x)(∀y)(∀z)(x ≤ X y → (y ≤ X z → x ≤ X z)), for each X ∈ S;
(∀y)(y ≤ X y), for each X ∈ S;
(∀y)(∀z)(y ≤ X z ∨ z ≤ X y), for each X ∈ S;
(∀x)(∀y)(Rϕ (x, y) → ϕ † (y)), for each ϕ ∈ L(P, S);
(∃x)ϕ † (x) → (∀x)(∃y)(∀z)(Rϕ (x, z) ↔ y = z), for each ϕ ∈ L(P, S);
(∀x)(ϕ(x) ↔ ψ(x)) → (∀x)(∀y)(Rϕ (x, y) ↔ Rψ (x, y)), for each ϕ, ψ ∈ L(P, S).
It is easy to verify (12.4) by direct inspection of the clauses of (12.6). In order to verify
(12.5), we argue as follows. Let A be a countable relational structure satisfying D. We
define a structure M = (W, s, u, t). First, let W = | A| and let t ( p) = Q Ap , for each p ∈ P.
Next, by (12.6)a–c, for each X ∈ S, ≤ XA is a countable linear preorder. It follows from the
universality of the rational numbers among countable linear orders that a utility function
u X may be chosen so that for all w, w ∈ W, w ≤ X w if and only if u X (w) ≤ u X (w ). For
each ϕ ∈ L(P, S) and w, w ∈ W, let s(w, ϕ[M]) = w if and only if Rϕ (w, w ). Finally,
let s(w, P) be an arbitrarily chosen element of P for any proposition P ⊆ W which is not
expressed by a sentence. It is easy to see that the structure M satisfies (12.5).
We now derive Theorem 6.5 from the compactness theorem for first-order logic. Let T
be a set of sentences of L(P, S) and suppose that every finite subset T ⊆ T is satisfiable. It
follows at once from (12.1) and (12.4) that for every finite T ⊆ T , {ϕ † (c) | ϕ ∈ T } ∪ D is
satisfiable (here c is a constant symbol). Therefore, by the Compactness and Löwenheim–
Skolem Theorems for first-order logic, there is a countable structure A such that A |
{ϕ † (c) | ϕ ∈ T } ∪ D. Hence, by (12.1) and (12.5), T is satisfiable.
2
BIBLIOGRAPHY
Andréka, H., Ryan, M., & Schobbens, P.-Y. (2002). Operators and laws for combining
preferential relations. Journal of Logic and Computation, 12, 12–53.
Blackburn, P., de Rijke, M., & Venema, Y. (2001). Modal Logic. New York, NY:
Cambridge University Press.
Dietrich, F., & List, C. (2009). A Reason-Based Theory of Rational Choice. Technical
Report, London School of Economics.
Gärdenfors, P. (1988). Knowledge in Flux: Modeling the Dynamics of Epistemic States.
Cambridge, MA: MIT Press.
Haidt, J. (2001). The emotional dog and its rational tail: A social intuitionist approach to
moral judgment. Psychological Review, 108, 814–834.
Johnson, P. E. (1988). Social Choice: Theory and Research. New York, NY: Sage
Publications.
Keeney, R. L., & Raiffa, H. (1993). Decisions with Multiple Objectives: Preferences and
Value Trade-Offs. Cambridge, UK: Cambridge University Press.
Lang, J., van der Torre, L., & Weydert, E. (2003). Hidden uncertainty in the logical
representation of desires. In Proceedings of Eighteenth International Joint Conference
on Artificial Intelligence (IJCAI03). San Francisco, CA: Morgan Kaufmann.
Levi, I. (1986). Hard Choices: Decision Making under Unresolved Conflict. Cambridge,
UK: Cambridge University Press.
Lewis, D. (1973). Counterfactuals. Cambridge, MA: Harvard University Press.
Liu, F. (2008). Changing for the better: Preference dynamics and agent diversity. PhD
Thesis, ILLC, University of Amsterdam.
PREFERENCE BASED ON REASONS
147
Messick. D. M. (1985). Social interdependence and decision making. In Wright, G., editor.
Behavioral Decision Making. New York, NY: Plenum, pp. 87–109.
Osherson, D., & Weinstein, S. (To appear). Modal logic for preference based on reasons.
In Tannen, V., editor. Lecture Notes in Computer Science. New York, NY: Springer.
Osherson, D., & Weinstein, S. (Forthcoming). Two approaches to the logic of preference.
Payne, J. W., & Puto, C. (1982). Adding asymmetrically dominated alternatives: Violations
of regularity and the similarity hypothesis. The Journal of Consumer Research, 9, 90–95.
Pettit, P. (2002). Rules, Reasons, and Norms. Oxford, UK: Oxford University Press.
Savage, L. J. (1954). The Foundations of Statistics. New York, NY: Wiley.
Sen, A. (1971). Choice functions and revealed preference. Review of Economic Studies,
38, 307–317.
Stalnaker, R. (1968). A theory of conditionals. In Rescher, N., editor. Studies in logical
theory. Oxford, UK: Blackwell.
Tentori, K., Osherson, D., Hasher, L., & May, C. (2001). Wisdom and aging: Irrational
preferences in college students but not older adults. Cognition, 81, B87–B96.
van Benthem, J., Girard, P. K., & Roy, O. (2009). Everything else being equal: A modal
logic for Ceteris Paribus preferences. Journal of Philosophical Logic, 38, 83–125.
van Benthem, J., & Liu, F. (2007). Dynamic logic of preference upgrade. Journal of
Applied Non-Classical Logic, 17, 157–182.
von Wright, G. H. (1963). The Logic of Preference. Edinburgh, UK: Edinburgh University
Press.
DEPARTMENT OF PSYCHOLOGY
PRINCETON UNIVERSITY
PRINCETON NJ 08540 USA
E-mail: [email protected]
DEPARTMENT OF PHILOSOPHY
UNIVERSITY OF PENNSYLVANIA
PHILADELPHIA PA 19104 USA
E-mail: [email protected]